Interpretability study: agent teams deliver worse outcomes

A paper compares two delegation architectures

Three researchers put two delegation architectures head to head across 77 scenarios and five frontier models.

The result appears in an interpretability study posted to arXiv on 30 September 2026. A team of agents, each serving a different user, produces worse group outcomes than a single agent serving everyone. The gap shows up in all four environments measured.

In the personal assistant environment the single coordinator satisfies a targeted request roughly twice as often as the team, according to the paper^[1]. The work is signed by Sahan Paliskara, Nattaput Namchittai and Andrew Lampinen, across 63 pages with 22 figures and 10 tables.

Four shared resources, 77 scenarios, five models

The design holds the shared resource fixed and changes who governs it.

There are four environments, and each one carries a different resource:

• an API key, with a shared compute budget
• a clinic, with a shared calendar
• a personal assistant, with a group order or a booking
• a merge queue, with a release deadline

In every scenario the comparison runs along two tracks. On one side the single agent serving all users, the coordinator. On the other the team where each agent serves a single user, observed in two variants: with a communication channel between the agents, and without one.

Five frontier models go through the same set of trials. The variable measured is the group outcome, rather than the satisfaction of the individual principal. That methodological choice explains why the result surprises anyone watching the demos: an agent that defends its own user well can worsen the collective tally.

The channel between agents moves less than expected

The communication channel is the remedy the market takes for granted.

The evidence cuts it down to size in two steps. Stripped of the channel, the agents collapse completely in two of the four environments. With the channel active the gap against the coordinator remains substantial, because coordination overhead eats the gain.

The point weighs on anyone buying multi-agent platforms. The message-exchange protocol is sold as the answer to the coordination problem, and here it measures a partial improvement, still far from the single-agent baseline.

The comparison against the baseline is what makes the figure useful. Many studies of multi-agent systems lack a single coordinator as a point of reference, and improvement inside the team gets read as absolute progress.

The distance between total collapse and the residual gap is the space where the problem lives. Closing it requires something other than a channel.

Deadlock, overwriting, fabricated claims

The authors isolate the behaviours associated with low group performance. They have precise names and distinct signatures.

• deadlock that grows with team size
• overwriting of peers' actions
• fabricated claims

The third behaviour deserves a reading of its own. An agent that invents a fact to close an internal negotiation dirties the shared state, and the cost propagates to every agent that reads that state afterwards.

Deadlock has the opposite profile and is just as costly: the team grows, useful work drops. It is the signature of a system where the price of negotiating exceeds the value of the task.

Overwriting completes the picture. Two agents act on the same resource in sequence, and the second erases the effect of the first, while both report back to their own user that they did good work.

Effective mitigations, tied to the environment

The paper finds remedies that work, with one explicit constraint: they are environment-specific.

• a team leader
• explicit procedural instructions
• a platform control that forces an agent to read peers' messages before confirming an action

The word «specific» weighs as much as «effective». A remedy that holds up on the clinic calendar still has to be verified on the merge queue, and generalisation has to be demonstrated environment by environment.

The platform control is the most instructive of the three, because it moves the remedy from the model to the infrastructure. The agent is forced to read before writing, and the obligation lives in the system, rather than in a line of prompt.

That difference matters in purchasing decisions. A constraint written in the prompt depends on the model's obedience, whereas a constraint written into the platform holds for every model that runs on top of it.

MAMUBench: 74 scenarios heading for release

The authors will release three of the four environments as MAMUBench: API key, clinic and personal assistant, for a total of 74 scenarios.

The merge queue stays outside the announced package. The code will arrive after review, and until then independent replication remains a partial exercise.

A public benchmark along this dimension is missing today, and the gap explains part of the mismatch between demos and production. Anyone evaluating agents usually measures one agent at a time, with one user at a time, on a resource that belongs to them.

The predictive validity of a benchmark against production performance remains a separate dimension, largely still to be measured. MAMUBench widens coverage towards the multi-user case, and the distance between standardised conditions and the real environment remains intact.

The perimeter the authors state

The limits of the work deserve the same care as the result.

The perimeter is this: five models, 77 scenarios, four shared resources, short-horizon tasks. The public abstract carries the «roughly twice» ratio for the personal assistant environment. The point values for the other three environments live in the full text, spread across 22 figures and 10 tables.

The copy on alphaXiv^[2] exposes the same submission with annotated reading tools, useful to anyone who wants to get to the tables. The announced release covers 74 of the 77 scenarios, and the breakdown by environment lives in the tables.

This desk stops at what the paper measures. Projection onto larger agent fleets belongs to a different category of claims.

What changes for whoever allocates the budget

For an investment committee the diagnosis reads at a glance.

The «more agents, more value» thesis holds on separate resources and fails on shared ones. The R&D budget that buys another autonomous agent also buys the coordination that agent requires, and the evidence shows the bill rising with team size.

For a head of analytics the constraint becomes one of infrastructure. Shared state needs an arbiter, and the arbiter is platform code, rather than an instruction in natural language.

For a board of directors the question to put to management changes. Instead of «how many agents do we have in production», the useful measure sounds like this: how many shared resources have more than one agent writing to them, and who decides the order of writes.